In (EC)DSA as per FIPS 186-4, the message to sign is first hashed. Imagine that we skip this hashing stage, instead put the message where the hash was, and constrain the size of message $h$ to the original hash's output width $N$ bits. The resulting scheme is vulnerable to (at least) these existential forgeries ($q$ is the multiplicative group order):
- any signature for message $0\le h<2^N-q$ is also valid for message $h+q$;
- a valid signature for messages $h=0$ and $h=q$ is easy to obtain: in DSA, $(r,s)$ with $r=s=y\bmod q$ does the trick, where $y$ is the public key; there's an analog with ECDSA, which $r=s=x_A\bmod q$, where $x_A$ is the $x$ coordinate of the public key, and $q=n$.
Are other attacks possible? In particular, does temporary access to a signing oracle allow a total break (key extraction or other mean to sign any message)?
The question is of direct interest when one wants to lower the communication overhead between a signing device and a host using it to (EC)DSA-sign large messages: can we fully offload the hash computation to the host? I believe that holds (and as noted in comments, that seems to be practice), but can we demonstrate that? If not, how can we safely offload most of that computation, without breaking standard conformance?